“Natural Disasters, Casualties and Power Laws: A ...personal.rhul.ac.uk/uhte/014/Natural...
Transcript of “Natural Disasters, Casualties and Power Laws: A ...personal.rhul.ac.uk/uhte/014/Natural...
“Natural Disasters, Casualties and Power Laws: A Comparative Analysis with Armed Conflict”
Óscar Becerra1, Neil Johnson2,3, Patrick Meier4,5, Jorge Restrepo3,6 and Michael Spagat7,3*
*In alphabetical order only
1Fedesarrollo, Bogotá, Colombia
2Department of Physics, University of Oxford, Oxford UK 3CERAC, Conflict Analysis Resource Center, Bogotá, Colombia
4The Fletcher School of Law and Diplomacy, Tufts University, Boston, US 5Center for Study of Civil War (CSCW), Peace Research Institute, Oslo (PRIO), Norway
6Department of Economics, Javeriana Univeristy, Bogotá, Colombia 7Department of Economics, Royal Holloway College, University of London, Egham, UK
Abstract Power-law relationships, relating events with magnitudes to their frequency, are common in natural disasters and violent conflict. Compared to many statistical distributions, power laws drop off more gradually, i.e. they have “fat tails”. Existing studies on natural disaster power laws are mostly confined to physical measurements, e.g., the Richter scale, and seldom cover casualty distributions. Drawing on the Center for Research on the Epidemiology of Disasters (CRED) International Disaster Database, 1980 to 2005, we find strong evidence for power laws in casualty distributions for all disasters combined, both globally and by continent except for North America and non-EU Europe. This finding is timely and gives useful guidance for disaster preparedness and response since natural catastrophes are increasing in frequency and affecting larger numbers of people. We also find that the slopes of the disaster casualty power laws are much smaller than those for modern wars and terrorism, raising an open question of how to explain the differences. We show that many standard risk quantification methods fail in the case of natural disasters. Paper prepared for presentation at the 2006 Annual Meeting of the American Political
Science Association (APSA), Philadelphia, PA, August 31-September 3, 2006. Contact: [email protected] or [email protected].
***PRE-CONFERENCE DRAFT ABSOLUTELY NOT FOR CIRCULATION***
Introduction
1
Power law distributions are frequently found in biological, physical and social systems
(Newman 2005, Buchanan 2000).1 For example, some natural disasters such as the size
versus frequency of earthquakes, forest fires and landslides follow a power law distribution.
Related research suggests that casualties in whole wars (Richardson 1948, 1960; Cederman
2003), global terrorist events (Clauset and Young 2005), and in war events (Johnson et al.
2005 & 2006) follow power laws.
Natural disasters are a major threat to human security. Purvis and Busby (2004, 68) reckon
that natural disasters affected roughly 188 million people per year between 1990-1999; an
order of 6 times the affected population for armed conflict over the same period.
Moreover, natural-disaster frequency is increasing over time (Emergency Disasters Data
Base or EM-DAT).
Knowledge of casualty distributions in natural disaster events can give critical support to
pre-disaster planning (Combs et al. 1999). A key component of contingency planning is the
identification of resources most likely to be in demand in a post-disaster environment when
disaster medical assistance teams (DMATs) and disaster mortuary operational response
teams (DMORTs) are deployed.2 Knowledge of the relative frequencies of disasters with
various casualty counts is vital for the rational stockpiling medical supplies, assembly of
rapid-response teams with the right numbers and skill mixes and insurance planning (among
other things).
In this paper we analyze natural disaster casualty distributions using EM-DAT’s Center for
Research on the Epidemiology of Disasters (CRED) International Disaster Database, 1980-
2005. We aggregate all disaster types and test for casualty power laws for individual
continents and for the whole world. 1 A key property of power laws is that their densities decline more gradually than those of many
common statistical distributions (such as the Guaßian or Normal distribution): the so-called “fat tail” property. In addition, both densities and distributions of power laws follow straight lines when plotted on log-log axes. See the appendix for some basic mathematics of power laws.
2 Of course, post-disaster casualty estimation is a critical component of any evolving response to a particular disaster that has already occurred (Sharma, 2001), but we do not consider this question in the present paper.
2
Here are our main findings. First, there is good evidence for power laws in disasters at the
global level. Second, power laws are well supported by continent, except for North America
and non-EU Europe. Third, the exact form (slope) of these power laws varies little from
case to case, i.e., the findings are robust to continent-by-continent disaggregation. Fourth,
natural disaster power laws are flatter than those arising from terrorism events on non-G7
targets (Clauset and Young, 2005) and modern war events (Johnson et al. 2005 & 2006).
This means that the relative frequencies of high-casualty events to low-casualty events are
higher for natural disasters than the corresponding ratios for non-G7 terrorism and modern
war, i.e., natural disaster casualty distributions have fatter tails. In fact, natural disaster
power laws are close to the power-law distributions that have been observed for casualties in
whole wars (Richardson 1948, 1960; Cederman 2003) and for terrorism on G7 targets
(Clauset and Young, 2005). Finally, because natural disaster risks follow a power law
distribution, we argue that commonly used variance-based risk measures are deeply
problematic in this context.
This paper is structured as follows: We first review the literature on casualty distributions in
natural disasters to provide context for the analysis that follows. Next we describe the data we
use and the methodology we apply in our study. We then present our findings, relate them to
the literature on war casualties, armed conflict and terrorism, and critique the standard risk
quantification approaches within the context of our finding. Finally, we consider the policy
implications for disaster risk management and outline possibilities for future research.
Literature Review: Casualty Distributions and Natural Disasters
Blong and Radford (1993), in a pioneering study, studied a record of disasters in the
Solomon Islands to assess the likely mixture of future events and contribute to the
development of risk and mitigation strategies. They used data developed by the Australian
Development Assistance Bureau on 209 disaster events, 27 of them occurring before 1900,
including cyclones, droughts, earthquakes, floods, storms, landslides, tsunamis and volcanic
eruptions. The authors’ plot of the cumulative distribution function for deaths in their
3
natural-hazard events on a log-log scale very much resembles a straight line. So the casualty
data for the Solomon Islands over a long period is suggestive of a power law, although the
authors do not mention or test for one.
Guzzetti (2000) studies a wide array of landslide data (not from CRED) going back very far
in time. He covers the periods 1410-1999 and 1950-1999 for Italy and similar but varying
periods for Canada, the Alps, Japan, Hong Kong and China.3 Through the inspection of log-
log plots Guzzetti (2000) finds that the cumulative frequency of landslide events plotted
against their casualty consequences “can be approximated by power laws”.
Jonkman (2005) uses the CRED dataset to study to study the distribution of killings in global
events, 1975-2001, disaggregating by disaster type including floods, droughts, earthquakes,
famines, windstorms and epidemics but not by continent as we do. For each disaster type
he plots the global frequency of events with N or more people killed against N on a log-log
scale. Only the curve for earthquake casualties resembles a power law and Jonkman does
not pursue this question. The near-absence of straight lines in these plots, together with our
own straight-line results for all disasters combined, raises a puzzle as to why power laws arise
globally and by continent for all disaster types but not for individual disaster types.
To summarize, we use completely different data compared to Blong and Radford (1993) and
Guzetti (2000). We use the same dataset as Jonkman (2005) but disaggregate by continent
rather than by disaster type. We find quite robust power-laws evidence whereas Jonkman
does not. We also contribute by introducing formal statistical testing for casualty power laws
into the disaster literature. We relate our finding to the literature on war casualties, armed
conflict and terrorism and we discuss the insurance implications of our finding.
Data and Methodology
3 There are, of course, significant quality issues arising in the use of such deep historical data. Nevertheless,
study is highly enterprising and interesting.
4
Our natural-disaster data is derived from the EM-DAT International Disaster Database
maintained by the Center for Research on Epidemiology of Disasters (CRED) in
cooperation with the United States Office for Foreign Disaster Assistance (OFDA). The
database defines a disaster as meeting at least one of the following criteria: 10 or more
deaths, 2000 or more affected people for droughts and famine or 100 or more for other
disasters, a government disaster declaration, or a plea for international assistance. We analyze
all disasters combined, 1980-2005, both globally and disaggregated by continent.
Disaster data is necessarily imperfect and the quality of the CRED data is known to vary
across years and disaster types. However, according to a quality assessment study by Guha-
Sapir and Below (2002), recent CRED data compares well in quality to proprietary datasets
and is the most reliable and comprehensive global dataset in the public domain. Dilley et al.
(2005) hesitate to draw on CRED data to identify total mortality risk, but are quite
comfortable using CRED to analyze relative risk. This is encouraging for our work since
power law analysis is specifically about estimating relative magnitudes since the latter also
focuses on relative magnitudes.
The ideas behind our estimation and testing procedures are as follows (details are in the
appendix). We estimate two parameters, minx and α , the minimum casualty level above
which the power law is supposed to hold, and the exponent of this estimated power law
respectively. We obtain these through an iterative procedure that starts with 1=minx ,
estimates the corresponding maximum-likelihood α and then repeats the procedures for
,...3,2=minx , settling on the ( )α,minx pair that gives the best estimated power-law fit to the
data according to the Kolmogorov-Smirnov statistic. We then test two hypotheses using
Monte Carlo methods to generate the test statistics: that the data is generated from the
estimated power-law curve and that the data is generated from the best-fitting lognormal
distribution. The lognormal is a commonly used fat-tailed distribution and, therefore, a
natural comparator. When we fail to reject the power-law hypothesis but do reject the
lognormal hypothesis we consider the evidence for a power law to be strong.
Results and Discussion
5
We summarize our findings for the whole world and by continent in Table 1. For each case
we give the number of disasters in the sample, the maximum number of casualties for these
events, our minx and α estimates, the number of observations greater than or equal to minx ,
and the p-values for the Kolmogorov-Smirnov tests of the power law (PL) and lognormal
(LN) hypotheses. In addition, Table 1 gives population, population density (per square
kilometer), and average GDP per capita adjusted for purchasing power parity.4
Casualties in Obs xmax \alpha xmin obs>xmin KS(Ho: PL)
KS(Ho: LN) Population Pop
densityGDP
per cap
World 7476 1800090 1.7304 177 995 0.434 <0.0001 6,525,486,603 49.81 11,479.20Americas 1808 1800090 1.7451 48 355 0.878 0.0180 893,496,906 22.96 12,198.00
North America 596 7060 2.1342 77 66 0.857 0.3408 331,672,307 17.71 34,540.00Center/South America 1212 1800090 1.6898 35 329 0.857 0.0100 561,824,599 27.83 9,715.56North America (with México) 730 39704 2.0877 86 89 0.932 0.3304 439,121,832 21.26 30,450.00C/South America (without México) 1078 1800090 1.6647 29 313 0.673 0.0084 454,375,074 24.88 9,709.09
Asia and Oceania 3135 277715 1.6934 215 527 0.745 <0.0001 3,920,701,817 101.37 9,190.14Africa 1519 300000 1.6631 39 432 0.163 <0.0001 910,849,725 30.56 3,916.07Europe 1011 61080 1.599 9 387 0.740 0.0008 800,438,155 33.91 22,706.12
EU countries 454 20000 1.7258 9 171 0.611 0.0008 454,809,846 118.73 26,173.91non EU countries 557 61080 1.7375 68 85 0.976 0.1664 345,628,309 17.48 19,638.46
Table 1. Estimation and testing results globally and by continent.
In most cases we find good evidence for power laws with α between 1.6 and 1.75. In
these other cases rejection of the lognormal is always with confidence above 98% and almost
always well over 99%. Africa is the only continent where we come anywhere close to
rejecting a power law at a standard significance level. So there is a strong robustness to the
findings.
The two exceptions are non-EU Europe where we cannot clearly reject the lognormal
distribution and North America where we are even further from being able to reject log-
normality the estimated α for the power law exceeds 2.1. One possible explanation for
these deviations is simply that these two cases have the smallest sample sizes. A second
possibility is that the different results are related to the fact that these two cases also have the
lowest population densities. It appears that we can rule out explanations based on per capita
income; although non-EU Europe and North America are relatively very rich so is EU
4 We took demographic indicators from the International Data Base of the U.S. Census Bureau
(http://www.census.gov/ipc/www/idbnew.html) and GDP per capita from the CIA World Factbook (http://www.cia.gov/cia/publications/factbook/). The GPD per capita values for the majority of countries are for 2005 but for a small group of countries there are estimates for earlier years.
6
Europe which behaves much like the poor countries in its natural disaster casualty
distribution.5
Figures 1 to 2 illustrate the power law estimation results. Figure 1a depicts our global result
while Figures 1b through 1e represent distributions by individual continent. Figures 2a
through 2f depict results by specific region. For each region, the empirical points are blue
dots and the estimated line is green and dotted. We mark the estimated minx with a vertical
line that is red and dotted and we display the estimated numbers for minx and α . The
Figures confirm the close fit of the estimates.
These power-law findings for casualties in natural disasters fit in nicely with established
results for whole wars, terrorist events and events in individual modern wars. Richardson
(1948, 1960) and Cederman (2003), treating an entire war similarly to the way we treat a
natural disaster in this paper find power laws with alphas, slightly higher but very close to
our estimated α ’s. Clauset and Young (2005) present similar results using terrorist attacks
on G7 targets as the basic unit of analysis. In contrast, Clauset and Young (2005) and
(Johnson et al. 2005 & 2006) find power laws with much higher α ’s, around 2.5, for
terrorist attacks on G7 targets and events in modern wars respectively. Some of these
papers have proposed models to explain their results. It is an open question whether a
model can be built to illuminate our natural disaster results.
5 We did not notice any other useful patterns in the relationships between the economic/demographic
indicators and our size distribution results.
7
Estimation result by total continents
1 10 50 100200 5001000 10000 1000000.0001
0.001
0.01
0.1
1
← xmin = 177
→ αML = 1.7304
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1a. World. α ≈ 1.73 and xmin = 177
Estimation results by continent
1 10 50 100 200 5001000 10000 1000000.0001
0.001
0.01
0.1
1
← xmin = 39
→ αML = 1.6631
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1 10 50 100 200 5001000 10000 1000000.0001
0.001
0.01
0.1
1
← xmin = 215
→ αML = 1.6934
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1b. Africa. α ≈ 1.66 and xmin = 39. 1c. Asia/Oceania. α ≈1.69 and xmin = 215.
1 10 50 100200 5001000 10000 1000000.0001
0.001
0.01
0.1
1
← xmin = 48
→ αML = 1.7451
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1 10 50 100 200 500 1000 100000.0001
0.001
0.01
0.1
1
← xmin = 9
→ αML = 1.5990
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1d. Americas. α ≈ 1.75 and xmin = 48. 1e. Europe. α ≈ 1.60 and xmin = 9. Figure 1: Estimation results. These graphs depict the casualty distribution across all disasters for all countries in the World, and disaggregated by continent.
8
Estimation results by selected regions
Americas regions
1 10 50 100 200 500 10000.0001
0.001
0.01
0.1
1
← xmin = 77
→ αML = 2.1342
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1 10 50 100200 5001000 10000 1000000.0001
0.001
0.01
0.1
1
← xmin = 35
→ αML = 1.6898
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
2a. North America. α ≈ 2.13 and xmin = 77. 2b. South/Center America. α ≈ 1.68 and xmin = 35.
1 10 50 100 200 500 1000 100000.0001
0.001
0.01
0.1
1
← xmin = 86
→ αML = 2.0877
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1 10 50 100200 5001000 10000 1000000.0001
0.001
0.01
0.1
1
← xmin = 29
→ αML = 1.6647
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
2c. North America (with Mexico). α ≈ 2.09 and xmin = 86.
2d. South/Center America (w/out Mexico). α ≈ 1.66 and xmin = 29.
Europe regions
1 10 50 100 200 500 1000 100000.0001
0.001
0.01
0.1
1
← xmin = 9
→ αML = 1.7258
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
1 10 50 100 200 500 1000 100000.0001
0.001
0.01
0.1
1
← xmin = 68
→ αML = 1.7375
P(X
≥ x
)
Severity of event, x - Number of people killed plus injured
2e. EU countries. α ≈ 1.73 and xmin = 9. 2f. Non-UE countries. α ≈ 1.73 and xmin = 68. Figure 2: Estimation results. These graphs depict the casualty distribution across all disasters for all countries disaggregated by specific regions.
9
Our findings have important implications for standard approaches to risk evaluation.
Jonkman et al. (2003) give a broad survey of quantitative methods for assessing and limiting
risks. There are a very wide range of risk measures in use. Generally, these are based either
on some combination of the mean and variance of the event distribution or they focus in
some way on the tail of the distribution. However, power law distributions have infinite
variance whenever α is below three and all our estimated α ’s almost always in the range of
1.6 to 1.75, i.e., well below 3. It is true that casualty distributions do not, in reality, extend all
the way to infinity; there are absolute limits to the number of people who can be killed in a
natural disaster event. Nevertheless, the infinite variance problem will manifest itself
empirically in high sensitivity of empirical variance measures to presence or absence of a
small number of big events. This problem is intuitive. Power laws have fat tails so high-
casualty events are fairly common. But the exact proportion of high-casualty events present
in a dataset at a given point in time will fluctuate rather erratically causing high variance in
any empirical estimate of the true variance of the underlying distribution. Thus, variance-
based risk measures are highly unstable.
A second standard approach to risk assessment base on tail behavior is the so-called FN
criterion. In the present context this would mean verifying that the probability of having x
casualties or more is always less than 2xC where C is a fixed constant. However, if x
follows a power law distribution with exponent α then this criterion reduces to 21
1xC
x<−α
for all x which is impossible when α is below 3 as it always is in our estimates.6 Again, the
problem here is intuitive. The criterion requires that event probabilities decline with their
severity at a rate that is inconsistent with the fat-tail property of a power law.
6 More concretely when 7.1=α , as in our natural disaster results, the inequality becomes 27.
1xC
x< which
of course will not hold for sufficiently high-casualty events.
10
Policy and Further Research
Our work suggests that there are strong regularities in casualty patterns in the broad
aggregate character of natural disasters as a whole. This pattern holds up quite well on most
continents. Moreover, the specific pattern is the fat-tailed power law. This means that we
should expect a steady stream of large-casualty natural disaster events. Large scale disasters
are not anomalies. They fall well within established patterns. Disaster risk-management
planning must take heed of this fact, which justifies investments being made to develop
more reliable disaster early warning systems for effective and early response.
The emergence of a power law distribution within an event list that mixes together many
disaster types suggests, to a significant degree, that all disasters with x casualties will resemble
one another and disaster response planning can proceed accordingly. Of course, this casualty
commonality can only be pushed so far. At some stage there must be more specific planning
for more specific types of disaster.
On the theoretical level our findings raise a question about whether aggregate patterns for all
disasters combined can be explained within a single framework. More importantly, can such
a theory provide concrete insights and recommendations to improve disaster management
and response?
Finally, standard methods for quantifying risks fail badly in the natural disaster context. It is
urgent to rethink these approaches in order to rationalize actual preparations.
11
Appendix
The probability density function for a power law in our context:
(1) p(x) = Cx−α
where x is the number of casualties in a given event and C is a constant such that the
probability of all possible outcomes ( x ) sums to 1.7 In equation (1), the probability that an
event results in a number of casualties equal to x is decreasing in x with the decline
depending on a “tail parameter” α where α is inversely related to the fatness of the tail. By
taking the logarithm of both sides, equation (1) is expressed as:
(2) xcxp log)(log α−=
That is, if the distribution of casualties belongs to the power-law class, xlog will be linear in
)(log xp . So a log-log graph )(xp against x is a straight line as is a log-log graph of
)( xXP > (1 minus the cumulative distribution function) against x .
Certain considerations must be taken into account when estimating the α coefficient. First,
the number of casualties is a variable of count data; it can take only integer values. Second,
empirical work shows that it is uncommon for the entire distribution to follow a power law.
Indeed, several researchers have found that only the ‘tail’ of the distribution, i.e. the right-
part of the distribution over a given value which we call minx , follows a power law.8 In
consequence, we use the power law distribution for discrete data above minx , as done by
Clauset and Young (2005). This distribution takes the form:
7 Formally, it means that C is a constant such that ∫
∞ − =1
1dxCx α . 8 See, for example, Price (1965), Shiode and Batty (2000), Clauset and Young (2005) and Newman (2005)
among others.
12
(3) minmin
xxx
xxp ≥=−
),()(
ας
α
where ∑∞
=−=
minxkmin kx αας ),( is the incomplete Riemann zeta function. Thus, our
estimation procedure identifies the optimal values of two parameters, minx and α . We
follow the procedure suggested by Clauset and Young (2005), which estimates the α
parameter using maximum likelihood procedures (Johnson and Kotz 1969, 240), and
identifies minx through minimization of the Kolmogorov-Smirnov (KS) goodness of fit test.
This tests the null hypothesis that the data belong to a specific distribution by comparing the
empirical )( XS and theoretical )( XF distributions:
(4) XX SFD −= max
While the optimal minx guarantees that this distance is the minimum, it does not necessarily
imply statistical significance. As a result, we implement a Monte Carlo approach to obtain
the p-value of this test. After obtaining the optimal minx and α values, we use the KS test
the hypothesis that the data follow a power law distribution, as well as the hypothesis that
the data follow the lognormal distribution, another popular right-skewed and fat-tailed
distribution.9
9 For further details above estimation procedure, see Johnson et al. (2005).
13
References Blong R.J., and D.A. Radford, “Deaths in Natural Hazards in the Solomon Islands,” published in Disasters (1993) 17: 1-11. Buchanan M., Ubiquity: The Science of History, (Weidenfeld & Nicholson, 2000). Burton I., Kates R.W., and White G.F., The Environment as a Hazard, Second Edition, (New York/London: Guilford Press, 1993). Cardona O., “The Need for Rethinking the Concepts of Vulnerability and Risk from a Holistic Perspective: A Necessary Review and Criticism for Effective Risk Management,” Bankoff, G., Frerks, G. and D. Hilhorst. (2004). Mapping Vulnerability: Disasters, Development & People (Earthscan, London). Cederman L., “Modeling the Size of Wars: From Billiard balls to Sandpiles,” Amer. Pol. Sci. rev., 97, 135-50 (2003). Clauset A. and Young M. “Scale invariance in Global Terrorism”, e-print physics/0502014 at http://xxx.lanl.gov (2005) Combs D.L., L.E. Quenemoen, R.G. Parrish, and J.H. Davis, “Assessing Disaster-attributed Mortality: Development and Application of a Definition and Classification Matrix.” International Journal of Epidemiology, vol. 28, pp. 1124-1129 (1999). Dilley M., Chen R., Deichman U., Lerner-Lam A., Arnold M. with J. Agwe, Buys, P., Kjekstad, B. L. and G. Yetman. (2005). “Natural Disaster Hotspots: A Global Risk Analysis,” Disaster Risk Management Series (The World Bank Hazard Management Unit, Washington DC). Espinosa-Aranda, J. M, Jimenez, A., Ibarrola, G., Alcantar, F., Aguilar A., Inostroza M., Maldonado S., and R. Higareda, “The Seismic Alert Systems in Mexico City and the School Prevention Program,” in Zschau, J. and A. Kueppers eds. (2002). Early Warning Systems for Natural Disaster Reduction (Springer Verlag, Berlin). Gilbert C., “Studying Disaster: Changes in the Main Conceptual Tools,” in Quarantelli, E. L. ed. (1998). What is a Disaster? Perspectives on the Question (Routledge London and New York). Guhar-Sapir, D. and R. Below. “Quality and accuracy of disaster data: A comparative analyse of 3 global data sets.” Working paper prepared for the Disaster Management facility, World Bank, Brussels CRED, 2002. Guzzetti, F. “Landslide Fatalities and the evaluation of landslide risk in Italy” Engineering Geology (2000) 58: 89-107 Heijmans A., “From Vulnerability to Empowerment,” in Bankoff, G., Frerks, G. and D. Hilhorst. (2004). Mapping Vulnerability: Disasters, Development & People (Earthscan, London).
14
Johnson N., Spagat M., Restrepo J., Bohorquez J., Suarez N., Restrepo E., and Zarama R., “From Old Wars to New Wars and Global Terrorism,” (2005) published on-line at: http://xxx.lanl.gov/abs/physics/0506213 Johnson N., Spagat M., Restrepo J., Becerra Ó., Bohorquez J., Suarez N., Restrepo E., and Zarama R., “Universal patterns underlying ongoing wars and terrorism,” (2006) published on-line at: http://xxx.lanl.gov/abs/physics/0605035. N.L. Johnson, and S. Kotz., Univariate discrete distributions. John Wiley & Sons, New York. (1992) Jonkman, S.N. “Global Perspectives on Loss of Human Life Caused by Floods,” Natural Hazards (2005) 34: 151-175; Kelman I., “Disaster Deaths,” access on July 8, 2006, available at: http://www.ilankelman.org/disasterdeaths.html Kirigia, J.M., L.G. Sambo, W. Aldis, and G.M. Mwabu. 2004. "Impact of disaster-related mortality on gross domestic product in the WHO African Region". BMC Emergency Medicine, vol. 4, no. 1, available at: http://www.biomedcentral.com/1471-227X/4/1 Newman M.E.J., “Power Laws, Pareto distribution and Zipf’s law,” Contemp. Phys. 46, N.5, 323-351 (2005). Pulwarty R., Broad K., and T. Finan, “El Nino Events, Forecasts and Decision-Making,” in Bankoff, G., Frerks, G. and D. Hilhorst. (2004). Mapping Vulnerability: Disasters, Development & People (Earthscan, London). Price D. J. “Networks of Scientific Papers”, Science 149, 510-15 (1965). Purvis, N. and Busby J., “The Security Implications of Climate Change for the UN System,” Environmental Change and Security Project, May 2004. Richardson L.F., “Variation of the Frequency of Fatal Quarrels with Magnitude, Amer. Stat. Assoc. 43, 523-46 (1948). Richardson L.F., Statistics of Deadly Quarrels, eds. Q. Wright and C.C. Lienau (Boxwood Press, Pitsburgh, 1960). Sharma, V. “Gujarat Earthquake: Some Emerging Issues”, Asian Disaster Management Vol. 7, No. 2 & 3, April-September 2001. Shiode N. and Batty M. “Distributions in Real and Virtual Worlds” Working paper series of the Centre for Advanced Spatial Analysis. 19 (2000) The Earth Institute, Risk Analysis Reports Over Half of World’s Population Exposed to One or More Major Natural Hazards. The Earth Institute, Columbia University, 2005.